Does scaling work for self-driving perception?

Autonomous driving perception systems face a unique challenge: fusing multiple sensor types (cameras, LiDAR, radar) while understanding 3D space. STELLAR tests whether the scaling approach that worked for large language models also works here. The team trained a Sparse Window Transformer on 50 million driving examples with up to 500M parameters and discovered clear scaling laws—performance improves predictably with model size and data. The result: new state-of-the-art on Waymo's open dataset, with substantial margins over prior systems.